Risk Stratification of Postoperative Enteral Feeding Intolerance Using Explainable Machine Learning in Oral Cancer Free Flap Reconstruction

Why It Matters

Accurate early prediction of FI enables clinicians to tailor nutritional strategies, potentially reducing complications and hospital stay for a high‑risk surgical cohort.

Key Takeaways

• Random Forest predicts feeding intolerance with AUC 0.889 using nine peri‑operative variables.
• High pre‑operative fasting glucose and low potassium markedly increase FI risk.
• Advanced tumor stage and longer operative time are strong metabolic‑inflammatory predictors.
• Lower Advanced Lung Cancer Inflammation Index signals higher postoperative feeding intolerance.
• Model uses routine data, enabling early, individualized nutritional interventions.

Pulse Analysis

Enteral nutrition is a cornerstone of recovery after extensive oral‑cancer surgery, yet feeding intolerance affects more than one‑third of patients undergoing free‑flap reconstruction. The condition not only compromises caloric delivery but also prolongs intensive‑care stays and raises the risk of infection. Traditional risk scores, derived from generic intensive‑care populations, fail to capture the unique metabolic stress, chronic inflammation, and surgical trauma inherent to head‑and‑neck oncologic procedures. Consequently, clinicians have lacked a reliable tool to identify patients who will struggle with early enteral feeding.

The study leveraged a large single‑center cohort to train seven machine‑learning algorithms on 35 pre‑operative variables, ultimately selecting a random‑forest model for its balance of discrimination (AUC 0.889) and calibration. Feature selection via LASSO distilled the predictor set to nine variables, with SHAP analysis revealing that elevated fasting glucose, low potassium, advanced tumor T stage, prolonged operative time, and a reduced Advanced Lung Cancer Inflammation Index (ALI) most strongly increased FI risk. These findings underscore a metabolic‑inflammatory axis: hyperglycemia and electrolyte disturbances impair gut motility, while a low ALI reflects combined nutritional depletion and systemic inflammation, both of which predispose to gastrointestinal dysfunction.

Clinically, the model’s reliance on routinely collected labs and operative data means risk stratification can occur before nutrition is initiated. High‑risk patients could receive proactive measures such as tighter glucose control, potassium supplementation, and a more gradual EN advancement schedule. Moreover, the interpretability offered by SHAP empowers clinicians to discuss specific risk factors with patients and multidisciplinary teams. While promising, the model requires external validation across diverse institutions before widespread adoption, and future research should explore integrating dynamic biomarkers or real‑time monitoring to further refine predictive accuracy.